Since ChatGPT became popular, various large models and generative AI technologies have emerged. Various industries such as medical care, finance, travel, consumer retail, and the Internet are looking to use generative methods AI technology empowers businessinnovation. However, from exploration to implementation, enterprises are applying generative AI There is still a threshold for technology, and the use of ready-made technical services will cause security concerns, such as business data leakage and other issues. For a time, many companies were in a dilemma. This article will talk about what companies are doing in the process of implementing generative AI applications.
-
From model selection to business security, the twists and turns of the birth of generative AI applications
In fact, the generative AI market has developed rapidly after more than a year. On the one hand, many large models and supporting services have emerged in the market. On the other hand, the problems faced by enterprises It is also increasing day by day.
First of all, starting from the selection of models, the dazzling large models are enough for manufacturers to drink. On May 28, the "China Artificial Intelligence Large Model Map Research Report" compiled by the China Institute of Scientific and Technological Information and the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology, jointly with relevant research institutions, was officially released. The report shows that my country's artificial intelligence large model map has a scale of more than 1 billion parameters. Nearly 80 large models have been released. In October, the "2023 China New Generation Artificial Intelligence Technology Industry Development Report" released by the China New Generation Artificial Intelligence Development Strategy Research Institute showed that the total number of domestic large models currently reaches 238. According to data from the Beijing Bureau of Economy and Information Technology, as of early October, Beijing had released 115 large models, including 12 general-purpose large models and 103 vertical large models.
According to this development trend, the "War of 100 Models" may soon escalate into the "War of 1,000 Models." It will be even more difficult for enterprises to choose large models. In specific application scenarios, enterprises need to weigh the balance between accuracy and performance, effectively compare models and find the best choice based on their preferred indicators, which requires deep data science expertise and consumes a lot of manpower and time. .
The selection of the model is just the beginning. After the model is determined, you need to do fine-tuning and training of the model based on your own business. At this step, the company's business data type needs to be adapted to the data type required for large model input. At the same time, the input data needs to be representative, diverse, consistent and reliable, so as to achieve better output. , which requires enterprises to have engineers who understand both business and large model technology to organize the data.
In addition, the fine-tuning of large models also requires a lot of computing power. It requires a lot of money and time to purchase and maintain hardware equipment or rent cloud services. At the same time, the infrastructure also requires long-term maintenance. As large model technology is an emerging technology, many companies do not have the corresponding talent reserves and experience, which is also a big pressure for enterprises.
After the problem of the model itself is solved, the enterprise will still face security risks. For example, many users are worried that if they use a large model, will their data be seen or even leaked by the model party? Will it lead to the leakage of sensitive information, or the generation of illegal content, etc.? Then, enterprises need to ensure that data will not be leaked or abused during transmission, storage and processing to avoid losses to business and reputation.
-
Amazon Cloud Technology helps enterprises securely build generative AI applications
Faced with various challenges, for many enterprises, the best choice may be to find themselves an AI assistant to assist in embedding AI capabilities in all aspects. The most common ones are large model platforms and services on the cloud.
On November 28, 2023, 2023 Amazon Cloud Technology re:Invent was grandly opened in Las Vegas, USA, and concluded successfully on December 2. Starting from December 12, 2023, the 2023 Amazon Cloud Technology re:Invent China City Tour will be launched in 10 major cities, covering 10 cities: Beijing, Shanghai, Guangzhou, Shenzhen, Chengdu, Qingdao, Nanjing, Xi'an, Hangzhou, and Changsha !
As an annual benchmark in the global cloud computing field, 2023 Amazon Cloud Technology re:Invent brings the latest product and technology releases, cutting-edge leader insights and global cloud computing to global cloud computing enthusiasts and builders best practices. In this year's re:Invent, a series of major releases became the focus of the conference, from serverless, data strategy in the generative AI era, chip and cloud base innovation, to Amazon Bedrock's major updates, enterprise-level generative AI applications (Amazon Q )'s new release, etc., a series of combination punches provide comprehensive protection for enterprises to implement generative AI applications.
Among several new announcements, one major update to the fully managed service stands out — Amazon Bedrock is releasing more model choices and new powerful features.
First, Amazon Bedrock integrates multiple industry-leading large language models and other models from AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon for easy access to users. Not only that, in order to help customers choose the one that is more suitable for them among the numerous models, Amazon Bedrock’s new function can help customers efficiently evaluate, compare, and select the model that best suits their application scenarios and business needs.
After determining the model, enterprises often need to adapt the model, or expand and optimize it, to maximize the value of the data. For these parts, the Amazon Bedrock knowledge base feature customizes model responses using contextual and relevant company data: Organizations looking to supplement existing models with proprietary data for more relevant and accurate responses. For model tuning, Cohere Command, Meta Llama 2, and Amazon Titan models in Amazon Bedrock support tuning, providing customers with more options for model customization. Anthropic Claude will also soon support tuning.
In addition, Amazon Bedrock also doubles as an agent that can automate, perform multi-step tasks using corporate systems and data sources. While models are effective at conducting conversations and creating new content, they can provide more value if they can perform complex actions such as solving problems and interacting with company systems to complete tasks (for example, booking a trip or ordering replacement parts). However, this requires custom integration of the model with company data sources, APIs, and internal and external systems. Developers must write code to coordinate the interaction between the model, the system, and the user so that the application can execute a series of API calls in a logical sequence. In order to connect a model with a data source, developers must deploy a RAG so that the model can tailor its response to the task. Finally, developers must configure and manage the necessary infrastructure and develop data security and privacy policies. These steps are time-consuming and require specialized knowledge, slowing down the development of generative AI applications. Now generally available, fully managed Amazon Bedrock agent capabilities enable generative AI applications to perform multi-step tasks across company systems and data sources.
Finally, at the level of interactive security. While many models use built-in controls to filter out inappropriate and harmful content, enterprises want to further customize interactions to ensure topics are always relevant to the business, comply with company policies, and adhere to the principles of “responsible AI.” For example, a bank might want to set up its online assistant to avoid looking up competitors, avoid providing investment advice, and limit harmful content. In addition, at the user's request, the user's personally identifiable information (PII) may be changed or anonymized to ensure security. Enterprises want to enforce key policies and rules in generated AI applications in a streamlined way to provide a question-and-answer user experience and enable safer use of the technology. Amazon Bedrock’s Guardrails feature, now in preview, enables customers to implement protections for generative AI applications. With Amazon Bedrock’s Guardrails capabilities, customers can implement protection measures across models based on application requirements and responsible AI policies. These applications are customized based on customer use cases and "responsible AI" principles, so this feature can enhance the security and privacy of user interactions.
At present, Amazon Cloud Technology has launched a variety of products to help enterprises build generative AI applications and alleviate AI "anxiety", including many partners in the technology industry. Sahir Azam, chief product officer at MongoDB, said: “More customers across industries are looking to leverage generative AI to build next-generation applications, but many are concerned about data privacy and the accuracy of output from AI-driven systems. To satisfy customers demand, we are using MongoDB Atlas as Amazon Bedrock’s knowledge base so that our joint customers can securely build generative AI applications using their operational data to create personalized experiences with the trust and accuracy end users expect. With this integration, customers can access industry-leading underlying models and use data processed by MongoDB Atlas Vector Search to create applications that deliver more relevant output in the right context. Leverage data built into the Amazon Bedrock Knowledge Base Privacy best practices, customers can save time spent on generative AI operations and focus more on technology deployment to deliver a more engaging end-user experience on Amazon cloud technology."
Of course, in addition to Amazon Bedrock, there are many major releases in 2023 Amazon Cloud Technology re:Inven, and in the following 2023 Amazon Cloud Technology re:Invent China event, Amazon Cloud Technology will also give an in-depth explanation of each new product and new function. .
Click here to view all the hot releases of Amazon Cloud Technology re:Invent 2023 in one link.